Evolutionary Structure Optimization of Convolutional Neural Networks for Deployment on Resource Limited Systems

Convolutional neural networks (CNNs) have achieved great success in various computer vision tasks. However, trial and error are still the most adopted way to design the structure of neural networks, which are time consuming. In this work, a population based evolutionary structure optimization approach of CNNs for deployment on resource limited systems is proposed. The method evolved several different kinds of individuals together in a way called natural selection. Evolutionary operators in conventional genetic algorithms are well defined based on structure design problem. Objectives of minimizing space cost or time cost of one deep neural network is considered in optimization process. Experiments on MNIST datasets show that the proposed method can evolve networks with state-of-the-art accuracy and have low storage or time cost, which give inspiration to hand-made structure by the obtained network structure.